On Friday 22 Mar 2024 at 14:24:35 (+0000), Quentin Perret wrote: > On Tuesday 19 Mar 2024 at 16:43:41 (+0000), Colton Lewis wrote: > > Add a KVM_CAP to control WFx (WFI or WFE) trapping based on scheduler > > runqueue depth. This is so they can be passed through if the runqueue > > is shallow or the CPU has support for direct interrupt injection. They > > may be always trapped by setting this value to 0. Technically this > > means traps will be cleared when the runqueue depth is 0, but that > > implies nothing is running anyway so there is no reason to care. The > > default value is 1 to preserve previous behavior before adding this > > option. > > I recently discovered that this was enabled by default, but it's not > obvious to me everyone will want this enabled, so I'm in favour of > figuring out a way to turn it off (in fact we might want to make this > feature opt in as the status quo used to be to always trap). > > There are a few potential issues I see with having this enabled: > > - a lone vcpu thread on a CPU will completely screw up the host > scheduler's load tracking metrics if the vCPU actually spends a > significant amount of time in WFI (the PELT signal will no longer > be a good proxy for "how much CPU time does this task need"); > > - the scheduler's decision will impact massively the behaviour of the > vcpu task itself. Co-scheduling a task with a vcpu task (or not) will > impact massively the perceived behaviour of the vcpu task in a way > that is entirely unpredictable to the scheduler; > > - while the above problems might be OK for some users, I don't think > this will always be true, e.g. when running on big.LITTLE systems the > above sounds nightmare-ish; > > - the guest spending long periods of time in WFI prevents the host from > being able to enter deeper idle states, which will impact power very > negatively; > > And probably a whole bunch of other things. > > > Think about his option as a threshold. The instruction will be trapped > > if the runqueue depth is higher than the threshold. > > So talking about the exact interface, I'm not sure exposing this to > userspace is really appropriate. The current rq depth is next to > impossible for userspace to control well. > > My gut feeling tells me we might want to gate all of this on > PREEMPT_FULL instead, since PREEMPT_FULL is pretty much a way to say > "I'm willing to give up scheduler tracking accuracy to gain throughput > when I've got a task running alone on a CPU". Thoughts? And obviously I meant s/PREEMPT_FULL/NOHZ_FULL, but hopefully that was clear :-)
